Using assembler trampolines instead of && fixes NMI IRET test.

There is another bug, however.  Returns to the same privilege level do
not pop SS:RSP on 32-bit, so the nested NMI underflowed the stack at "s".
This is fixed in the new code.  The new code doesn't set up SS:RSP and
instead leaves some space for the nested NMI handler on the alternate
stack.  The old stack pointer is kept and restored when the nested
handler returns.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>